<p>Recent discussions of Artificial Intelligence (AI) safety often invoke the orthogonality thesis—the idea that intelligence and final goals can be independent variables, allowing for the possibility of radically alien AI agents. This paper challenges this view by examining the relationship between language, intentionality, and rational agency against the background of how we now think about Large Language Models (LLMs). Drawing on Robert Brandom’s normative inferentialism and accounts of LLMs as role-playing dialogue agents, we argue that the linguistic competence demonstrated by LLMs suggests significant constraints on their artificial agency. We contend that participation in human discursive practices, which has an inherent normative dimension, may bind specific artificial agents to forms of rationality that are continuous with human reasoning. By analysing recent work on LLM linguistic behaviour, we demonstrate how these systems engage with human normative practices in ways that complicate AI sceptics’ dismissal of LLM capabilities and AI safety theorists’ speculations about future AI agents. We defend Brandom’s approach against charges of insufficient naturalism that could be put forward by the AI safety theorist camp while arguing that his account of language and rationality as inherently social practices provides crucial insights into the nature and potential limits of artificial agency. Our analysis suggests that the development of language-capable AI systems may be constrained by the normative structure of human discursive practices in ways that bind specific quadrants of the space of possible artificial minds.</p>